<Comment deleted by user>
I read a part of this blog. And also learned about some arguments. It seems that the critique, is about how to combine elements that are otherwise learn in isoltaion via hindsight.
I also agree that that the superposition work has rarely been done, and that the foreseeing problem in light of the unknown size from the learning point of view of where a game might be going given their experience set, is neglected.
And I think that is linked to the rigidity of features that carry their chunk of currency by themselves. And there is not superpositon concern when they happen in foresight, either on current position, or the mind's eye imagination, for many walks of chess. I could assume that the long opening sequence I learned, validated by theory to have good evaluation and even its official plan ideas, are all I need to know to succeed, I would not know really all the other paths, and might stick to the knowledge by procuration I am clinging to, and winning with so far.
I the lowest level features did not have to be called out only when they are valuable in hindsight, then the foreseeing problem could be adress in a more rational way. Explict position context information would argue (hunch or theory or both, I don,t see why one should prevent the other, if the prerequisite more neutral features of board are made more combinable.
I would refer to the book Point count chess, or PCC. It is the only place where I have seen the question even adress of the superpostion problem using exisint chess theory named features. And they had to contort them to make a rational first attempt at a table of such featues to actual commensurable values of evaluation nature.. so they can be guage somewhat naturally. They still have had to make per game examples modulation, as the raw concepts included might not have be conceived as ensemble of features. But they tried, and that aspect has gone completely unseen by reviewers..
The other chess theory that has had to try that, was the classical SF evaluatoin function, at the leafs, and only a proportion of all position of the large set of positions or chess wilderness (we might want to talk about one day, btw, as some elephant in the chess culture room). Yet their evaluation function can't be only used in hindsight.
So they have been betting (or used to) on hand-crafted theory that had to have ubiquitous applicatino and explicit dependencies, and be applied in the foresight problem direction. Since exhaustive search engine program was born on the bet that a weak leaf evalauatoni position was good enough the more exhaustive the search was (and deep), it took A0 and LC0 new blood, and better quality evaluation for any position (or also unspecified large set of position or wilderness scope, yet, a bit more controls as part of the learning desing conceptualization priori to implementation, but still, it has the RL dilemna, that exploration has to compromise with winning. The more ones knows how to win, the less enclined to to test far from previous winning hypotheses... (I am condensing more than one paper here. possibly cutting corners).
But now SF is using something closer to what I might be talking about. The let the NN do the evaluation job, out of more neutral features, algerbraic fucntions of the position information. While in the hand crafted, those features would be black holes in linear combinations, one parameter or few parameter per features. now it is a flexible function basis that is doing the high dimensional callcus, somehow. with clunky engine tounrmament optmization (not as controlled learning from outcome as a0 and LC0 are doing before even playing in tournaments). But with blurred vision, the engine familiy of which SF is the top horse, might be learning also from outcomes, but its leanring batches are its tournament versions. Its dev training, is of same nature. But they call it testing.. and is more aking to hyperameter tuning. It is not as clear cut as for A0 and LC0.. how the chess board feedback is improving the engine... from experience throughthe Elos.. (which are a statistics one might related to outcome information). but it does not directly or with formalism control the parameters space search. sorry for the techinicalityies.. i am trying to build an analogy, or used those machine experiements as the most rational models of learning to contribute to this topic here.
Howevere, in the master NN training for the NNue, and with the actuall constraint that the NN is trained to fit not chess game outcome information directly in a mathematically controlled way as fulll NN from a0 or Lc0 have been implementing, but to fit its simple evaluation in exhaustive but partial tree search context, a moderate depth further (A sort of information feedback loop from itself, on leaf taken as new root of deeper searches).
Still, that is building a theory of learning for the NNue, learning SF engine search at depth, not chess wilderness, or unspecified covering of that. It has not been on the radar of anyone AFAIK. to consider that variable of what is that large set.... A0 an dLC0.. have actually tried as part of their prior. but then they have to win. so only first batch is asking that question.. after that priority is to win.
main point.. out of this too much information, but I want to give some sense of the depth of those experiements as source of thinking possibly relevant to a pragmatic approach, main point is that the input space of the NN is not yet valuablatoin. it is more like the seeing.. the difficult foresight omelette of features to combien and evaulate as an orchestra of features partitions all having their contribution interacting with all other featyures, is left to a well formalized method of function sapce global optimizatno, that will have maximal proper generalization over whatever simple evalulation and agressive forward pruing modern day exhaustive engine descendiant have been considering the large set of position (implicitely, as there is no work on asking that question rationally).
we want SF NNue input space. it could help this thread discussion.. We need not the implementation code, and the quantization prowess, but the chess-land language version. I know there are now capable layman communicators on the SF team.
Thank you for your explanation. In fact, I've stumbled over the question of "metarules" before, and I agree that it can't be answered in general.
Still, I challenge your premises.
My view may be confined not only to my level of chess ability, but also to certain habits and preferences of thinking. Nevertheless, I'll try to explain in the same formal way as you introduced your axioms.
First, it's irrelevant to me if you can come up with additional reasonable principles or if one is written in an influential book by a world champion. It's only relevant if I a) know and b) have chosen to heed that principle.
There are principles I learnt about but chose not to employ beause they don't seem helpful, or I don't fully understand them, or I'm in doubt as to how to employ them (or probably all of the above).
So in fact, there *IS* an official, approved list of principles *I* apply - it's right here in my brain. ;-)
Only principles that have proven useful make the list.
And therefore, your Premise 2 is wrong for me and my list of principles:
In a given position, only one or even no principle at all may apply.
Obviously, with only one principle, you won't need a meta-rule.
So having a list of principles that at least sometimes apply alone, and in that case are good *on average*, improves one's decision making on average.
Next, you state that there are no meta-principles or meta-rules.
In fact, there are: For instance: "In a closed position, delay in deployment is not as bad as in open positions."
or "In a rook endgame, having the active rook is more important than having as many or more pawns."
Nobody said that a metarule needs to be a simple priority list.
In fact, we take decisions in real life with some "fuzzy" logic:
We identify important and less important arguments, and weigh them.
And in fact, I claim that *everybody* applies principles (inlcuding computers), even if not naming them.
I'll give you an example:
I guess we all agree, again computers included, that being a queen down is bad *on average*.
Isn't that a principle?
Indeed, it is!
And of course, it doesn't hold universally. It's easy to come up with examples where the weaker side can force mate, or stalemate, or repetition; or promotion with winning the opponent's queen, or with any of the above.
Now here is how I evaluate positions with a queen down: Either I see one of the cases listed above can be forced - or the position is *bad*.
In fact, I suppose that very many players will agree on that.
Of course, there are bound to be examples where it is wrong. However, when it *is* wrong, then probably beacuse we don't _see_ the right variant(s). Knowing that there is a path to victory or draw doesn't help if we don't find that one.
So this is a simple example of a final evaluation rule, based on a chess principle, which helps most players.
I don't see a reason to ditch it.
Do you?
Addendum:
There's even a simpler example.
Consider a position where you may consider a move M1.
You notice that after M1, your opponent has a forced way to checkmate you.
Question: How do you evaluate M1?
I suppose: Very bad. In fact, it ends up in a position that can't be worse for you.
How do you come up with that evaluation?
Well, it is a chess _rule_, in the literary sense of the word.
And there is no doubt about the meta-rules on evaluation here:
Mate trumps everything. It doesn't matter if you have the better pawn structure after you lost.
Again, a simple example of a rule, and its relationship to other rules and principles (wich you called metarules).
@MrMatt96 said in #10:
> This series has been articulating a lot of things I've never been able to verbalize very well, but which have been rattling around my brain for a long time. Eagerly awaiting the next installment!
>
> I do want to push back gently against a couple components of this part (mostly places where I felt like I was only 80% agreeing with you, rather than 100%.)
>
> I agree with the core premise that most of what we do at the chessboard is intuitive, and that "principles" are typically just explanations for intuitive judgements. I think it logically follows that experience is the best teacher; that our chess intuition is developed by seeing and playing out a variety of positions, rather than rote learning and applying a list of principles. But I do think principles unavoidably play at least some role in the development of intuition, even beyond just the early beginner-level.
>
> Principles are unavoidably imperfect proxies for concrete truths about any given position, or about chess knowledge more generally, but they do let us codify and clarify broad approximations of chess knowledge, which expedites the learning process. Take the concept of an outside passed pawn, for example. We could start with a basic principle, something like, "outside passed pawns are generally very good in the endgame, often winning." This would obviously, as you point out, have to undergo basically an infinite number of revisions and sub-principles, some of them contradictory or needlessly complicated: "Outside passed pawns are better in knight endgames than bishop endgames;" "Outside passed pawns are worse than protected passed pawns;" "Outside passed pawns are sufficient to win in rook endgames, but only if you can put your rook behind it, and only if it's an extra pawn, and only if there are enough pawns remaining on the other side of the board, and only if the king position is favorable, and only if the opponent lacks sufficient counterplay..." etc. However! The core principle is still a unit of chess knowledge which can be taught, learned, and eventually subsumed into a broader intuitive framework.
>
> I learned about outside passed pawns from my dad when I was maybe 11 years old, and I thought they were a wonderfully exciting thing. There's a very silly game of mine from about this time where I spent several moves in the early middlegame working to prepare the possibility of an outside passed pawn, then relaxed, content in the certainty my position would eventually win itself. Of course, I promptly got steamrolled everywhere else on the board and lost a largely non-competitive game. But I learned some things about outside passed pawns! I learned some things which could, after the fact, be reduced into some sort of verbalized principle, one which would have been of limited usefulness. My learning was largely intuitive, done at the board, but it was jump-started by the conceptualization that there was this thing called an "outside passed pawn" which could be very powerful in given situations.
>
> In the thousands of games I've played since then, I've refined my intuitive feeling around outside passed pawns. Many of those feelings could certainly be distilled into some sorts of principles (and I think there's a valid case to be made for the usefulness of principles simply as a refined and verbalized approximation of part of some more complex chess knowledge), but most of the learning has gone directly into building my intuition, without the in-between step of having to define explicit principles.
>
> That said, it's not at all clear to me that I would have been able to develop this intuition nearly as effectively without the starting point of a designed principle which I could use as a guidepost, nor, to a lesser extent, without the help of curated examples codifying particular sub-principles and exceptions which I likely would not have understood or even necessarily encountered on my own.
>
> Basically, I think you're right that principles are, on average, fairly useless at revealing deep truths about a given position, but I do think they oughtn't to be shortchanged as an important part of the chessic learning process.
Thank you for your support and your well-thought-out comment! I think you could be very well right regarding the role of principles in the development of chess intuition. It's plausible that the awareness of a defined concept, such as an outside passed pawn, could indeed expedite the process of developing intuition in positions featuring that concept.
This could perhaps be because explicitly defining a principle brings more attention to that aspect of the position, potentially accelerating the development of intuitive feel in similar contexts/positions featuring that concept.
However, it’s challenging to determine definitively whether your intuition around passed pawns would have developed differently without the explicit knowledge of the concept and its associated principles. The theory supporting the idea that explicit principles can speed up intuition development may depend on the notion that they highlight specific aspects of the game, which in turn may facilitate more focused 'intuitive absorption'.
This then raises further questions. Would a simple definition suffice to bring the necessary attention to an aspect of a position, or is a more comprehensive principle required? Additionally, if principles do indeed catalyze the development of intuition, to what extent do they help, at what chess level, and is the effort 'worth it' in the grand scheme of a player’s development? There may also be negative effects resulting from a focus on principles, and then weighing up the positives and negatives becomes another consideration.
Ultimately, when dealing with such a nebulous and complex skill as intuition, which often operates like a black box, it is difficult to deduce definitive answers through pure reason alone.
> For instance: "In a closed position, delay in deployment is not as bad as in open positions."
> or "In a rook endgame, having the active rook is more important than having as many or more pawns."
Ahum!, did you not just fix how they are usually communicated. Not as relatively as here, and often (other text in post above) not with the probabilistic divulging.
Also thanks, I had not seen that last one before. I will consider it as a hypothesis of understanding, as one possible aspect to ponder on my own experience, but it makes some sense, although many pawns can be structured to annoy the .hit out of one rook. If they are not aligned on rook mobility-subspace predilection. In a chain, one might need to protect only one pawn from the rook. I guess there might be another probabilistic statement, things happening more often, than such a chain. But still, I get it. Probably, the pawn structure (or placement) would still have a buffet zone for the rook (rank 2 or rank 7).
One can fix a bunch of such orphaned from proper nuances made explicit "principles", but the problem the blog raises, (that I read so far), is still there. That isolated argument (called principle) is only one isolated argument, when facing the foresight question, i.e. when playing.
Most of what we call principle are borne out of hindsight.. Even the book authors are doing it, at first creation of chess theory from their own intuition. The words don't fall on their laps. The concept might have from board experience and previous illustrious theory they might be critiquing or challenging with trepidation, in some pamphlet (Tarrash, Nimzo, theater).
I like that there are on commensurable posts in the discussion. I can make my usual ramblings, and feel less alone. Thanks to all here, and the author for inciting this, in blog. Now: where was I in my multiple pointers of forward reading in the blog series. I wonder what I missed in blog #3. I only recall axiom1, then I went back to forward reading on #1 and #2.
@NochEinSpieler said in #14:
Thank you for your detailed comment. I have read through your response a couple of times but am still confused by certain parts, so feel free to correct any misinterpretations.
> First, it's irrelevant to me if you can come up with additional reasonable principles or if one is written in an influential book by a world champion. It's only relevant if I a) know and b) have chosen to heed that principle.
> There are principles I learnt about but chose not to employ beause they don't seem helpful, or I don't fully understand them, or I'm in doubt as to how to employ them (or probably all of the above).
> So in fact, there *IS* an official, approved list of principles *I* apply - it's right here in my brain. ;-)
> Only principles that have proven useful make the list.
To this I would say that premise 2 is not an individual-specific statement. In other words, it's not saying that there are a number of applicable principles *that the player knows of* that are in conflict with each other. Principles in chess stem from the idea that there is an inherent logic within chess, and that certain rational patterns repeat themselves. For example, that pushing pawns in front of your castled king is bad. These principles and patterns are true independent of whether you know them or not. If premise 2 was relative to a single individual player, then of course it could be possible that this player is only familiar with a single principle, and hence will 'apply' that same principle in every single position.
> So having a list of principles that at least sometimes apply alone, and in that case are good *on average*, improves one's decision making on average.
Even here, it's unclear how a list of principles will causally impact your decision making. You need to outline the precise mechanism by which a principle will actually influence what move you decide to play. For example, consider your extra queen example:
> And in fact, I claim that *everybody* applies principles (inlcuding computers), even if not naming them.
> I'll give you an example:
> I guess we all agree, again computers included, that being a queen down is bad *on average*.
> Isn't that a principle?
> Indeed, it is!
> And of course, it doesn't hold universally. It's easy to come up with examples where the weaker side can force mate, or stalemate, or repetition; or promotion with winning the opponent's queen, or with any of the above.
>
As I've argued, making a move that can be justified by a principle after the fact doesn't mean that you applied that principle. Rather, for the principle to be applied, it must have influenced you to change your mind as to what move to make after you considered the principle. If I see a variation where I simply lose my queen for nothing, I don't need to recall or consider that principle it to immediately and intuitively know that move is bad.
I'll also quickly note here, just to be clear, that my argument does not say anything about computer evaluation, as that operates in a completely different way from chess player evaluation.
>
> Next, you state that there are no meta-principles or meta-rules.
> In fact, there are: For instance: "In a closed position, delay in deployment is not as bad as in open positions."
> or "In a rook endgame, having the active rook is more important than having as many or more pawns."
> Nobody said that a metarule needs to be a simple priority list.
> In fact, we take decisions in real life with some "fuzzy" logic:
> We identify important and less important arguments, and weigh them.
>
I don't believe I said that there can be *no* meta-rules at all, and if I did imply that at any point then it was likely an unintentional mistake, as it's clearly untrue. For example, take a specific position, position A, where two principles are competing but one ends in the player getting mated. Then a meta-rule would be that in position A, we should prefer principle 1 over principle 2, as to not get mated. Rather, my claim with respect to the existence of meta-rules is that there cannot exist some comprehensive set of meta-rules.
> So this is a simple example of a final evaluation rule, based on a chess principle, which helps most players.
> I don't see a reason to ditch it.
> Do you?
Again, I fail to see how you've demonstrated the process by which a principle helps most players. To do so, you would need to provide examples where consideration of principles *actively caused* a change in initial move preference, without falling back on some vague and nebulous intuition-driven process that some other commentors have suggested.
In my opinion principles are a type of patterns. Most of the patterns that we call principles can be put in the form Pi: [(certain characteristics of the position) and (certain characteristics of the best/good/bad move/plan)]. Let's denote it by Pi: [CPi and CBi]. Each Pi has a robustness percentage, which is the percentage of times CBi holds given that CPi holds; and each Pi has a level of difficulty of learning, depending on factors such as how hard it is to recognize the presence of CPi and/or CBi and generally inversely correlated with both the relative frequency of CPi and Pi's robustness. A player may learn some of these patterns and apply them during a game either consciously or subconsciously. A player may only talk about these patterns or explain these patterns if they have acquired conscious awareness of them. For other purposes I agree that acquiring conscious awareness of the patterns is not strictly necessary - and becoming aware of them can have both positive and negative impacts on skill level.
I agree that there is no absolute order of importance of principles applicable in any position. However, there are higher order (meta) patterns of the form MPi: [(all the relevant characteristics of the position) and (relative (partial) ordering of principles in terms of importance)]. These meta-patterns can also be learned and applied and understood. However, they usually have a higher level of difficulty - for example one reason for this is that the relative frequency of a certain set of (all the relevant characteristics of the position) is usually quite low. What you call intuition in this context is I think an - often subconscious - understanding and application of these higher order patterns. These higher order patterns are expected to be far from 100% robust, so our intuition can often be misleading.
All of these different types of patterns seem to be part of a "pattern space" together with mating patterns, tactical patterns, patterns concerning the relative strength of pieces, pawn structure, king safety etc. None of these patterns are explicitly built into the game, rather they emerge in a pattern recognition machine like a human brain (although at least some of them could be arrived at through deductive rather than inductive means as well) through repeated application of the rules of chess (playing/calculating chess positions) and then can be used as imperfect shortcuts to help us find our way through the huge realm of possibilities.
I'm looking forward to seeing how this point of view compares to yours. Great series so far!
@DailyInsanity said in #16:
> Thank you for your support and your well-thought-out comment! I think you could be very well right regarding the role of principles in the development of chess intuition. It's plausible that the awareness of a defined concept, such as an outside passed pawn, could indeed expedite the process of developing intuition in positions featuring that concept.
Exactly.. So glad this crowd of thinkers can uncover things that could have been long ago. If we lifted our heads sooner.
But the wrapper of delivery of such misnamed "principles" would benefit from making that part of the theory of communication (basis of any theory of teaching aware of the theory of learning or improving if one must be some tangible) explicit in the presentation or sharing to learner (or anxious, possibly jumping up and down, impatient improvers).
This does not mean the exercise you are doing is not worth doing to the max. It might help clarify the extent of the world of non-unicorns out there. I do like my wilderness, I live in an asphalt and concrete town nowadays, but was raised with lots of greenery when I had some luck of birth still not wasted. So, wilderness is the holy grail, it is now exotic, and is my sense of "out there".
And then so, now that this shared hunch, can put a bridge between knocking some sense back into existence of intuition as dominating pretty much all our grandiloquent attempts at sharing truths, one pamphlet or other scripture at a time, we might consider not throwing the baby with the bathwater, (by now I mean after your series has been digested by the many of us, including my humble patzerness self, of undetermined performer level). But no hurry. Better follow your wisdom in being clear at the reasoning level. If what I perceive as such in indeed one of your intentions.